speridlabs releases "Idol," a model that generates 3DGS from a single image, on Hugging Face

What happened

speridlabs has released Idol, a model that generates 3D Gaussian Splatting (3DGS) from a single image of an object, on Hugging Face. The input can be a rendered image, a generated image, or a photo. The model removes the background and crops the image automatically, then generates up to 1,048,576 Gaussians. It says the output includes sides that are not visible in the image.

How it works

Inference combines an image encoder, a diffusion-based generator, and decoders. The list in the model card gives about 370M parameters for the DiT (denoiser), about 50M for the octree decoder, and about 240M for the Gaussian decoder.

DINOv3 ViT-H+/16 and the FLUX.2 VAE are used to read the input, and BiRefNet is used for background removal. DINOv3 and FLUX.2-dev are gated models on Hugging Face, and you must agree to their terms of use.

The output is a standard 3DGS .ply. The model also writes out the input image as the model saw it (after background removal and cropping) and the estimated camera as JSON.

Choose --mode according to the type of input. synthetic is for renders and generated images of a single object, and crops so that the object fills 1/1.2 of the frame. real is for photos, and crops so that the object fills 1/1.7.

Speed-up is on by default. It uses torch.compile, batches guided and unguided inference into a single batch, and uses cuDNN attention and TF32. The authors say this gives about a 1.7x speedup without changing quality. The first image is slower because of compilation.

By the numbers

Why it matters

Because 3DGS can be made from a single image, possible uses include drafting small props and background assets for games and video production. The output is a standard .ply, so it is said to open in existing 3DGS viewers.

At 7.2GB of VRAM, it should be possible to try on a local GPU rather than a workstation. However, with the same settings the RTX 3080 takes 46.2 seconds and the RTX 4090 takes 19.1 seconds, so processing time varies widely by GPU.

For records that require dimensions and fidelity, such as cultural heritage or infrastructure, it is not a substitute for measured data, because unseen sides are filled in by generation. The source material does not mention such uses either.

Examples

• In the model card's benchmark, Idol is compared with TripoSplat, TRELLIS, and LGM. Input views are evaluated with PSNR, SSIM, LPIPS, and IoU, and 7 novel views with DINOv2, CLIP, and sharpness distance. One H100 was used per model.
• In the table, Idol shows the best value on every metric except LGM's PSNR. LGM's PSNR is 12.46, higher than Idol's, and the author notes that "even a blurry result tends to score high on mean squared error."

How to try it

1. With your Hugging Face account, agree to the terms of use for facebook/dinov3-vith16plus-pretrain-lvd1689m and black-forest-labs/FLUX.2-dev, then log in with hf auth login (or HF_TOKEN).
2. Clone the GitHub repository listed in the model card (speridlabs/idol-sdk) and install it with pip install -e .
3. For a photo, run python -m idol.sample --image photo.jpg --mode real. For a folder of rendered images, run it with --mode synthetic --out outputs.
4. Open the output .ply in any 3DGS viewer. First check with the default 262,144 Gaussians, and increase --num_gaussians if you need more detail. If you prioritize speed, reduce --steps, though detail is said to drop.

Caveats and open questions

• All benchmarks and processing times are values published by the model's authors. Third-party verification and a paper cannot be confirmed from the source material.
• The source states it "can run in under 2 seconds," but the conditions are not given in the material. With the table's setting of 50 steps and 262,144 Gaussians, even the fastest B300 takes 3.4 seconds.
• License terms and whether commercial use is allowed cannot be confirmed from the material obtained. Check the model card and repository before use.
• Use requires agreeing to the terms of two gated models, and an NVIDIA GPU and CUDA are required.
• Sides not visible in the input are filled in by generation, so there is no guarantee they match the real object. The model is intended for single-object input, not for reconstructing whole scenes.
• The benchmark measures similarity between the input image and the model's output. The contents of the image set used for evaluation cannot be confirmed from the material.

Glossary

3D Gaussian Splatting (3DGS)

A method that represents 3D space with many semi-transparent ellipsoids (Gaussians) and renders it quickly.

DiT

Diffusion Transformer. A diffusion model, which generates data step by step from noise, built with a Transformer.

classifier-free guidance (CFG)

A mechanism that adjusts how strongly the generated result follows the input condition (here, the image).

PSNR / SSIM / LPIPS

Metrics for measuring image similarity. Higher is better for PSNR and SSIM, and lower is better for LPIPS.

IoU

A metric of how much two regions overlap. Here it appears to measure how well the object silhouettes match.

Sources

Primary sources (publisher, paper, repository)

• speridlabs/idol · Hugging FaceHugging Face(speridlabs)（huggingface.co）・en

The X post that surfaced this story

• View @aisearchio's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original